A Deep Learning Method for Ocean Front Extraction in Remote Sensing Imagery

Abstract

Oceanfront is a typical and important sea surface feature that is reported to be associated with marine ecosystems and can be used as a reference for locating fishing grounds. Frontal zone extraction is often performed using a gradient threshold to classify image pixels and the result can sometimes contain too many spikes and become chaotic, leading to a negative effect for visual interpretation. And most conventional methods that focus on extracting the ridges of fronts struggle with false fronts due to imperfect data. Also, choosing appropriate thresholds for them is another dilemma, which sometimes leads to too many frontal ridges in unwanted areas or too little than needed in the region of interest. To meet the needs for visual interpretation and automatic front detection in significant frontal areas, a novel method based on deep learning is proposed in this letter. In this method, a deep learning model with U-Net architecture was designed to detect and locate significant frontal zones in grayscale sea surface temperature (SST) images. Then, an area threshold was adopted to filter the output of the model to improve the result. To demonstrate the effect of the proposed method, it was applied to an example SST image. The results show that the proposed method can not only merge messy fronts but also capture the overall patterns of frontal zones and work with conventional methods to get a better frontal ridge extraction result.